Zoo 2.3, now with 100% more Claude

I’ve released Zoo 2.3 after living with Zoo 2.2 for a while.

I gotta say: Zoo 2.2 has been a huge step forward for my autonomous coding. I routinely run more agents in parallel, and trust it with more complex tasks. I’ve started fully using up weekly rate limits on two Pro 200 subscriptions, too.

Here’s what’s new:

Claude Code is back

Claude runs the same workflow as Codex now, which is a significant step up in quality compared to Zoo 1. Codex and Claude skills are exact duplicates of each other, so maintaining them is easy.

Codex on GPT 5.5 xhigh still does better than Claude on Opus 4.7 xhigh and even Opus 4.7 max. Did not have a chance to experiment with Opus 4.8 yet.

Final reports

For when a very informative “Done” is not a satisfying end to a multi-hour agent run. A new Zoo Report skill runs at the end of a Zoo workflow and gives you a useful summary of what the agents did, crucially including screenshots for all UI changes.

Proposals tame uber-reviews

Uber-review has a problem: if you ask enough reviewers to inspect code, some will flag issues way outside the current scope:

• this subsystem has the wrong abstraction
• this permission model is suspect
• this UI primitive is duplicated everywhere
• this rollout story is missing

All true, possibly. Also, congrats, we were fixing one checkbox.

Now Zoo has options:

• small local cleanup stays within the (sub)task
• mundane broader cleanup becomes a separate subtask/commit
• bigger cross-cutting refactors become proposal files requiring human approval

Zoo Rebase & Zoo Push

If you run multiple Zoos in parallel, getting changes pushed used to be a real chore. Now:

1. Zoo rebases on top of upstream after finishing changes. Rebase is smart enough to rerun tests if there were incoming changes or merge conflicts. This Zoo Rebase skill can also be invoked manually.

2. You can manually run Zoo Push, which does Zoo Rebase, tries to push, rebases again if necessary, and so on until the push succeeds.

(Note: Zoo Rebase isn’t called Zoo Pull to avoid unfortunate Pull/Push mistakes when autocompleting the name.)

Zoo Ensure Safe Deploy

For that extra bit of snake oil you need to convince yourself to deploy without reading the code^†.

^† Practice not actually recommended by the author.

Tests pass, reviews pass, the code looks fine, and your stomach still says “lol no”. What else can you do?

Well you could try burning all your remaining tokens on getting a 100% assurance from your favorite coding agent. Run /goal (Codex) or /loop (Claude) with Zoo Ensure Safe Deploy inside.

By default it reviews unpushed commits, but you can point it at something else.

The skill is a variant of:

Ensure safe deployment of unpushed commits. We have a bunch of changes sitting here that we're too chicken to deploy. So I want you to de-risk this deployment. You need to dig into the changes, and (if necessary) production configurations and data, and carefully compare code behavior before and after this change, until you can give me firm 100% guarantee these changes are safe to deploy, or can produce failing tests for any REAL-WORLD problems.

…but said in a lot more words, and with a focus on expending significant effort.

Upgrading

Point your agent at github.com/andreyvit/zoo and ask it to install/upgrade your Zoo.